Enhanced Active Learning of Convolutional Neural Networks: A Case Study for Defect Classification in the Semiconductor Industry

With the advent of high performance computing and scientific advancement, deep convolutional neural networks (CNN) have already been established as the best candidate for image classification tasks. A decisive requirement for successful deployment of CNN models is the vast amount of annotated images, which usually is a costly and quite tedious task, especially within an industrial environment. To address this deployment barrier, we propose an enhanced active learning framework of a CNN model with a compressed architecture for chip defect classification in semiconductor wafers. Our framework unfolds in two main steps and is performed in an iterative manner. First, a subset of the most informative samples is queried based on uncertainty estimation. Second, spatial metadata of the queried images are utilized for a density-based clustering in order to discard noisy instances and to keep only those ones that constitute systematic defect patterns in the wafer. Finally, a reduced and more representative subset of images are passed for labelling, thus minimizing the manual labour of the process engineer. In each iteration, the performance of the CNN model is considerably improved, as only those images are labeled that will help the model to better generalize. We validate the effectiveness of our framework using real data from running processes of a semiconductor manufacturer.